Re: Additional compiler barrier required in sched_preempt_enable_no_resched?

[Vikram Mulukutla]

On 5/13/2016 7:58 AM, Peter Zijlstra wrote:
> On Thu, May 12, 2016 at 11:39:47PM -0700, Vikram Mulukutla wrote:
> > Hi,
> >
> > I came across a piece of engineering code that looked like:
> >
> > preempt_disable();
> > /* --cut, lots of code-- */
> > preempt_enable_no_resched();
> > put_user()
> > preempt_disable();
> >
> > (If you wish to seriously question the usage of the preempt API in this
> > manner, I unfortunately have no comment since I didn't write the code.)
>
> I'm with Thomas here, that's broken and should not be done.

Ok. I did in fact zero in on this code by replacing each instance of 
preempt_enable_no_resched with preempt_enable one by one (there were 
several uses in the driver). I will ask the original developer to 
consider using preempt_enable.

> > This particular block of code was causing lockups and crashes on a certain
> > ARM64 device. The generated assembly revealed that the compiler was simply
> > optimizing out the increment and decrement of the preempt count, allowing
> > put_user to run without preemption enabled, causing all sorts of badness.
> > Since put_user doesn't actually access the preempt count and translates to
> > just a few instructions without any branching, I suppose that the compiler
> > figured it was OK to optimize.
> >
> > The immediate solution is to add a compiler barrier to the code above, but
> > should sched_preempt_enable_no_resched have an additional compiler barrier
> > after (has one before already) the preempt-count decrement to prevent this
> > sort of thing?
>
> I think the below would be sufficient; IIRC the compiler may not combine
> or elide volatile operations.
>
> ---
>   include/asm-generic/preempt.h | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/include/asm-generic/preempt.h b/include/asm-generic/preempt.h
> index 5d8ffa3e6f8c..c1cde3577551 100644
> --- a/include/asm-generic/preempt.h
> +++ b/include/asm-generic/preempt.h
> @@ -7,10 +7,10 @@
>
>   static __always_inline int preempt_count(void)
>   {
> -	return current_thread_info()->preempt_count;
> +	return READ_ONCE(current_thread_info()->preempt_count);
>   }
>
> -static __always_inline int *preempt_count_ptr(void)
> +static __always_inline volatile int *preempt_count_ptr(void)
>   {
>   	return &current_thread_info()->preempt_count;
>   }

Thanks Peter, this patch worked for me. The compiler no longer optimizes 
out the increment/decrement of the preempt_count.

Thanks,
Vikram